Introduction

Moneypenny is the intelligence and governance core for AI agents. It provides structured memory, policy-governed execution, explainable audit, and portable state — all in a single SQLite file per agent.

It can run as a standalone agent runtime, or plug into existing runtimes — Claude Code, Cortex Code CLI, OpenClaw — as the memory, policy, and audit substrate. One command to connect: mp setup claude-code.

Why Moneypenny

Most agent frameworks solve orchestration. Moneypenny solves what happens between orchestration calls:

• What does the agent remember from last week?
• Is this action allowed by policy?
• Why was that tool call denied?
• How do two agents share knowledge without a central server?
• Can I back up my agent by copying one file?

Everything Moneypenny can do is exposed as an MCP server, so you can interact with it through natural language in any MCP-compatible client — Claude Code, Cortex Code CLI, OpenClaw, or any tool that speaks the Model Context Protocol. No CLI memorization required.

Four Pillars

Intelligence System

Structured memory with adaptive compression, hybrid retrieval (FTS + vector), graph-linked facts, and automatic extraction from conversations.

Governance & Security

Policy checks on every action, denial-aware control flow, auditable decisions, secret redaction, and encryption at rest.

Correctness Core

Single ACID turn semantics, database-as-runtime execution, and rollback guarantees that prevent partial state.

Scale & Portability

CRDT sync between agents, scoped knowledge sharing, one-file portability across devices and environments.

How It Works

The database is the runtime. Inference, memory, search, sync, policy, and tools live inside the same transactional boundary. The orchestrator is a thin loop; the intelligence sits between the DB and the LLM.

User message (CLI, MCP, or HTTP)
  → Context assembly (facts + knowledge + history + skills)
  → Policy evaluation
  → LLM generation
  → Tool execution (policy-gated, audited)
  → Fact extraction
  → Response

Every step operates on the same SQLite file. No external services required. The entire surface area is available over MCP, so every operation — memory writes, fact searches, policy checks, ingestion — can be triggered by natural-language requests in an MCP-connected chat client.

Use It Your Way

Moneypenny exposes three equal access paths. Pick whichever fits your workflow, or mix and match.

MCP / Natural Language

Connect Moneypenny with one command (mp setup claude-code, mp setup cortex, or mp setup openclaw). Then just ask:

Ingest the README from our docs repo.

What do you know about rate limits?

List all active policies.

The client translates your request into the right MCP tool call automatically.

CLI

The mp binary gives you direct access from any terminal:

mp ingest path/to/document.md
mp facts search "rate limit"
mp policies list

HTTP API

Start the gateway and hit the REST endpoints:

mp start
curl http://localhost:8080/v1/agents/main/facts?q=rate+limit

Useful for integrations, CI pipelines, and custom dashboards.

Built On

Seven SQLite extensions, statically linked into one binary:

Component	Role
sqlite-ai	On-device LLM inference, embeddings, chat (GGUF)
sqlite-vector	Vector search, SIMD, quantization, 6 distance metrics
sqlite-memory	Persistent agent memory, hybrid search
sqlite-rag	Hybrid search, RRF, multi-format docs
sqlite-sync	CRDT offline-first sync
sqlite-js	User-defined JS functions and aggregates

Next Steps

Quickstart Get running in under 5 minutes.

Core Concepts Understand agents, memory, policy, and the data model.

CLI Reference Every command, flag, and subcommand.

Architecture How the canonical operation layer works.